Why AI won't save your promotion ROI

The evolution of AI in TPM and RGM

Trade promotion management has gone through three distinct phases of technological maturity. Understanding where we are now - and what each phase actually requires - is essential for knowing why AI deployments succeed or fail.

Each phase is technically more capable than the last. And yet the 60% number hasn't moved. The reason is that every phase inherits the same underlying problem.

What AI models in TPO actually need

A machine learning model trained to predict promotional lift needs a specific data architecture to function. This is not a soft requirement - it is a hard technical constraint. The model's predictive accuracy is a direct function of the quality and completeness of its training data.

SKU-level transaction history

Lift models require clean historical sell-out or sell-in data at the SKU/retailer/week level, typically a minimum of two to three years of history to capture seasonal patterns and account for promotional cycling. Data must be consistent in granularity - mixing channel-level and account-level data without normalization produces systematically biased predictions. Missing weeks, restated baselines, or reclassified SKUs all degrade model performance in ways that are difficult to detect without careful feature validation.

Condition records and promotional mechanics

The model needs to understand what was actually executed, not what was planned. In most SAP environments, promotional conditions live in VK11/VK12 condition records - but the executed discount often differs from the planned condition due to retrospective adjustments, partial deliveries, or manual overrides in customer finance. If the model is trained on planned rather than actual conditions, its lift estimates are systematically off for high-deduction accounts.

Master data integrity

This is the most underestimated technical dependency. AI models in RGM typically operate across product hierarchies: brand, sub-brand, format, SKU, EAN. If the product hierarchy is inconsistent - for example, a packaging change treated as a new SKU without continuity mapping, or EAN codes that vary between internal systems and retailer data - the model cannot correctly attribute historical performance. It treats a reformulated product as a new launch, losing two years of relevant training signal.

The same applies to customer hierarchies. A national account split into regional buying groups mid-year, without retroactive adjustment of historical data, creates a structural discontinuity in the training set. The model sees a volume collapse at the national account level and a volume spike at the regional level - and cannot distinguish this from an actual promotional effect.

"Master data is not an IT concern. It is a model accuracy concern. Every inconsistency in the product or customer hierarchy is a source of bias in every downstream AI prediction."

The feature engineering dependency

Even with clean underlying data, TPO models require deliberate feature engineering to capture the dynamics of trade promotion. Standard features include:

• Baseline separation - isolating the non-promoted baseline volume from the promoted lift. This requires consistent flagging of promotional weeks in the historical data, which depends on accurate condition record history.
• Cannibalization signals - tracking volume shifts to adjacent SKUs within the same category during promotional periods. This requires a complete and consistent product hierarchy to define what counts as "adjacent."
• Retailer-specific price elasticity - different accounts respond differently to the same promotional mechanic. This requires sufficient transaction history at the account/SKU level to estimate elasticity with statistical confidence - typically 50+ promotional events per account per category.
• Forward buy detection - identifying when promotional volume represents genuine consumer offtake versus retailer stock-building. This requires sell-out data, not just sell-in, and consistent retailer reporting standards.

Each of these features fails gracefully when master data is inconsistent. The model trains, produces predictions, and passes validation checks - but the predictions are subtly wrong in ways that only emerge during post-event analysis, months after the promotional investment was made.

Why the organizational setup is a technical problem

The standard framing is that poor data quality is a process problem - a matter of training, accountability, and workflows. That framing is incomplete. In the context of AI-enhanced TPM, poor data quality is a model accuracy problem with direct financial consequences.

Sales organizations are typically structured around account management profiles: results-driven, commercially oriented, focused on the customer relationship. The maintenance of condition records, the upkeep of product hierarchies, the reconciliation of planned versus actual promotional conditions - these tasks require a different cognitive orientation. They require someone who treats data consistency as a deliverable, not as overhead.

The practical implication for AI deployment is concrete: if the people responsible for system input change jobs every 18 months, take shortcuts during quarter-end, or lack the incentive to retroactively correct condition records, the training data for your lift model degrades continuously. The model's accuracy erodes silently. Post-event ROI looks worse than expected. The AI recommendation is blamed. The real cause is the data pipeline upstream of it.

The implementation sequence that works

Based on implementations across FMCG organizations, the sequence that consistently produces results is:

• Audit master data first - before selecting a TPO platform, run a completeness and consistency audit on product hierarchies, customer hierarchies, and historical condition records. Quantify the gaps. This step is unsexy and takes two to four weeks. It also determines whether your AI investment will produce a return.
• Assign data ownership explicitly - a dedicated data analyst or trade analyst role, with master data maintenance as a primary KPI. Not the account manager. Not a shared responsibility. One owner per data domain.
• Establish actuals reconciliation as a closing process - at month end, planned promotional conditions should be reconciled against actual deductions from customer finance. This reconciliation feeds directly into the model's training data for the next cycle.
• Start with descriptive before predictive - a clean Power BI view showing actual vs. planned promotional performance at the SKU/retailer level is more valuable than a miscalibrated lift model. Build confidence in the data layer before adding the AI layer on top.
• Then automate - once the data foundation is stable and the descriptive layer is trusted, the move to AI-driven optimization is straightforward. The models have clean training data. The predictions are accurate. The flywheel turns: better predictions improve planning, better planning improves execution data, better execution data improves the next prediction cycle.

The opportunity when it works

When the data foundation is in place, the AI layer in TPM and RGM delivers measurably. ML-based lift models with clean training data consistently improve promotional ROI by 10-15% versus rule-based planning (TPO platform benchmarks, 2025). Real-time optimization in agentic RGM systems can reduce trade spend waste by identifying underperforming promotions during execution rather than after.

The 60% failure rate is not a ceiling. It is the current cost of deploying capable AI on a broken data foundation. Fix the foundation, and the number moves.